Lightweight forest fire detection method based on DDG-YOLO

Abstract

The invention relates to the technical field of forest fire detection, and particularly discloses a lightweight forest fire detection method based on DDG-YOLO. The method comprises the following steps: firstly, collecting and constructing an image data set of a forest fire scene, and preprocessing the data set; then, on the basis of a YOLOv8 network model, a DDG-YOLO improved model is constructed by introducing a densely fused bidirectional feature pyramid network DF-BiFPN, a dynamic up-sampling operator Dysample and a shared convolution detection head GN-SCDH fused with packet normalization; and finally, training the DDG-YOLO improved model, selecting the model with the best performance in the training process to detect the forest fire image, and generating a detection result. Experiments show that the DDG-YOLO improved model significantly reduces the complexity of the model while improving the detection precision, and is suitable for real-time monitoring of forest fire in a resource-constrained environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of forest fire detection, and in particular relates to a lightweight forest fire detection method based on DDG-YOLO. Background Art

[0002] Forest fires not only damage the ecological environment and threaten human life, but also cause enormous economic losses. Forest fires spread rapidly and are difficult to effectively control in a short period of time. Therefore, developing an efficient and fast-response forest fire detection system is crucial.

[0003] In vast forest areas, traditional forest fire detection methods, such as those relying on manual patrols and watchtowers, have significant limitations, primarily limited detection range, slow response times, and high costs. In contrast, drone-based forest fire detection methods offer significant advantages. Drones enable large-scale, highly efficient monitoring and can operate effectively in complex terrain and harsh environments.

[0004] The key to drone-based forest fire detection methods lies in machine vision object detection technology, which is primarily categorized into two-stage and one-stage detection algorithms. Two-stage detection algorithms, such as those in the R-CNN series, first generate candidate regions and then predict the object category and location for each region. While these algorithms offer high accuracy, they suffer from poor real-time performance and require significant computing resources. In contrast, one-stage algorithms, such as those in the SSD and YOLO series, while slightly less accurate, offer a simpler structure and faster inference speed, making them more suitable for real-time applications.

[0005] In forest fire detection, drone-captured images feature complex backgrounds and widely varying object scales. Furthermore, existing methods often struggle to find an ideal balance between detection accuracy and processing speed. Summary of the Invention

[0006] In view of this, the main object of the present invention is to provide a lightweight forest fire detection method based on DDG-YOLO.

[0007] In order to achieve the above object, the present invention provides a lightweight forest fire detection method based on DDG-YOLO, comprising the following steps:

[0008] S1: Construct an image dataset of forest fire scenes and preprocess the dataset;

[0009] S2: Reconstruct the neck network structure based on the YOLOv8 network model: design a densely fused bidirectional feature pyramid network DF-BiFPN to optimize the neck network feature fusion, and use the dynamic upsampler Dysample to replace the traditional upsampling module Upsample;

[0010] S3: Improved detection head structure: Introducing the shared convolutional detection head GN-SCDH with fusion group normalization to replace the original detection head and build an improved DDG-YOLO model;

[0011] S4: Train the DDG-YOLO improved model, select the model with the best performance during training to detect forest fire images, and generate detection results.

[0012] The following is a detailed description of the above steps based on the specific content:

[0013] Optionally, the specific implementation method of step S1 includes the following steps:

[0014] S11: Collect an image dataset of forest fire scenes and add confusing scenes such as sunset glow, clouds, and fog to the dataset;

[0015] S12: Use LabelImg software to label the dataset, and the output labeling file is a txt file in YOLO format;

[0016] S13: The dataset is randomly divided into training set, test set and validation set in a ratio of 7:2:1;

[0017] S14: Adjust the dataset image resolution to 640×640.

[0018] Optionally, the neck network structure reconstructed based on the YOLOv8 network model in step S2 is constructed by the following method:

[0019] S21: Using the weighted feature fusion idea of the bidirectional feature pyramid network BiFPN, design the DF-BiFPN structure;

[0020] S22: Dysample is used instead of the traditional Upsample to optimize the upsampling process of the feature map.

[0021] Optionally, the GN-SCDH in step S3 is constructed by the following method:

[0022] S31: Replace the group normalization of the convolution module with batch normalization to construct CGS convolution. The three feature maps output by the neck network are each passed through a 1×1 CGS convolution to unify the number of channels.

[0023] S32: Input the result into the shared convolution layer for feature extraction. This layer contains two 3×3 CGS convolutions in series.

[0024] S33: Process features through shared classification convolutional layers and regression convolutional layers, and introduce a Scale layer in the regression branch for feature scaling.

[0025] Optionally, the specific implementation method of step S4 includes the following steps:

[0026] S41: Set the hyperparameters for model training, perform model training, and obtain the optimal model for target detection;

[0027] S42: Use the test set to verify the detection effect of the optimal model for object detection, compare it with the original model based on the evaluation indicators, and verify the improvement effect of each module through ablation experiments;

[0028] S43: Inputting the forest fire image to be detected into the optimal target detection model for detection to obtain the target detection result.

[0029] Optionally, the specific implementation method of step S21 is as follows:

[0030] In traditional feature fusion methods, feature maps of different scales are weighted with fixed weights, resulting in insufficient feature adaptability and the inability to dynamically adjust the importance of features at each scale according to the characteristics of the detection target.

[0031] BiFPN dynamically adjusts the importance of feature maps of different scales in the fusion process by introducing learnable weights, thereby more efficiently fusing multi-scale features. Its weighted feature fusion adopts a fast normalization method, as shown in formula (1). This method scales the weights to the range of 0 to 1 and uses the ReLU activation function to ensure that w i ≥0, and set ε=0.0001 to avoid numerical instability.

[0032]

[0033] Among them, O is the output feature, I i represents the input feature, ε is a numerical stability constant, and w i With w j Represents different weight learning parameters.

[0034] While retaining the weighted feature fusion mechanism of BiFPN, DF-BiFPN adopts a three-stage feature fusion strategy (top-down, bottom-up, and top-down) to construct a densely fused bidirectional feature pyramid network. Specifically, the structure first performs channel adjustment on the feature maps of the P3, P4, and P5 layers output by the backbone network through a 1×1 convolution, while simultaneously downsampling the feature map of the P2 layer through a 3×3 convolution. In the first top-down feature fusion stage, high-level features are gradually downsampled through 3×3 convolutions, and the downsampled results are weightedly fused with features at adjacent scales. Next, the network maintains the original bidirectional feature pyramid architecture of YOLOv8, and through the bottom-up and second top-down feature fusion processes, it ultimately constructs a pyramid structure that can fully integrate multi-scale feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a lightweight forest fire detection method based on DDG-YOLO of the present invention.

[0036] Figure 2 This is a structural diagram of the DDG-YOLO improved model of a lightweight forest fire detection method based on DDG-YOLO in the present invention.

[0037] Figure 3 Schematic diagram of the GN-SCDH structure of a lightweight forest fire detection method based on DDG-YOLO in the present invention.

[0038] Figure 4 Schematic diagram of the neck network structure constructed based on DF-BiFPN of a lightweight forest fire detection method based on DDG-YOLO of the present invention.

[0039] Figure 5 This is a detection effect diagram of a lightweight forest fire detection method based on DDG-YOLO in the present invention. DETAILED DESCRIPTION

[0040] In order to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0041] The embodiment of the present invention discloses a lightweight forest fire detection method based on DDG-YOLO, such as Figure 1 As shown, the method includes the following steps:

[0042] S1: Construct an image dataset of forest fire scenes and preprocess the dataset;

[0043] S2: Reconstruct the neck network structure based on the YOLOv8 network model: design a densely fused bidirectional feature pyramid network DF-BiFPN to optimize the neck network feature fusion, and use the dynamic upsampler Dysample to replace the traditional upsampling module Upsample;

[0044] S3: Improved detection head structure: Introducing the shared convolution detection head GN-SCDH with fusion group normalization to replace the original detection head and build the DDG-YOLO improved model. The specific structure is as follows Figure 2 As shown;

[0045] S4: Train the DDG-YOLO improved model, select the model with the best performance during training to detect forest fire images, and generate detection results.

[0046] The following is a detailed description of the above steps based on the specific content:

[0047] The specific implementation method of step S1 includes the following steps:

[0048] S11: Collect an image dataset of forest fire scenes and add confusing scenes such as sunset glow, clouds, and fog to the dataset;

[0049] S12: Use LabelImg software to label the dataset, and the output labeling file is a txt file in YOLO format;

[0050] S13: The dataset is randomly divided into training set, test set and validation set in a ratio of 7:2:1;

[0051] S14: Adjust the dataset image resolution to 640×640.

[0052] The specific implementation method of step S2 of reconstructing the neck network structure based on the YOLOv8 network model includes the following steps:

[0053] S21: Using the weighted feature fusion idea of the bidirectional feature pyramid network BiFPN, design the DF-BiFPN structure;

[0054] S22: Dysample is used instead of the traditional Upsample to optimize the upsampling process of the feature map.

[0055] The GN-SCDH of step S3 is constructed by the following method:

[0056] S31: Replace the group normalization of the convolution module with batch normalization to construct CGS convolution. The three feature maps output by the neck network are each passed through a 1×1 CGS convolution to unify the number of channels.

[0057] S32: Input the result into the shared convolution layer for feature extraction. This layer contains two 3×3 CGS convolutions in series.

[0058] S33: Process features through shared classification convolution layer and regression convolution layer, and introduce Scale layer in regression branch for feature scaling. The specific structure of GN-SCDH is as follows: Figure 3 shown.

[0059] The specific implementation method of step S4 includes the following steps:

[0060] S41: Set the hyperparameters for model training, perform model training, and obtain the optimal model for target detection;

[0061] S42: Use the test set to verify the detection effect of the optimal model for object detection, compare it with the original model based on the evaluation indicators, and verify the improvement effect of each module through ablation experiments;

[0062] S43: Inputting the forest fire image to be detected into the optimal target detection model for detection to obtain the target detection result.

[0063] Furthermore, the DF-BiFPN in step S21 is characterized by:

[0064] In traditional feature fusion methods, feature maps of different scales are weighted with fixed weights, resulting in insufficient feature adaptability and the inability to dynamically adjust the importance of features at each scale according to the characteristics of the detection target.

[0065] BiFPN dynamically adjusts the importance of feature maps of different scales in the fusion process by introducing learnable weights, thereby more efficiently fusing multi-scale features. Its weighted feature fusion adopts a fast normalization method, as shown in formula (1). This method scales the weights to the range of 0 to 1 and uses the ReLU activation function to ensure that w i ≥0, and set ε=0.0001 to avoid numerical instability.

[0066]

[0067] Among them, O is the output feature, I i represents the input feature, ε is a numerical stability constant, and w i With w j Represents different weight learning parameters.

[0068] On the basis of retaining the BiFPN weighted feature fusion mechanism, DF-BiFPN adopts a three-stage feature fusion strategy of "top-down-bottom-up-top-down" to construct a densely fused bidirectional feature pyramid network. Specifically, the structure first adjusts the channels of the P3, P4, and P5 layer feature maps output by the backbone network through a 1×1 convolution, and at the same time downsamples the P2 layer feature map through a 3×3 convolution. In the first top-down feature fusion stage, high-level features are gradually downsampled through 3×3 convolutions, and the downsampled results are weightedly fused with adjacent scale features. Then, the network maintains the original bidirectional feature pyramid architecture of YOLOv8, and through the bottom-up and second top-down feature fusion processes, it finally constructs a pyramid structure that can fully integrate multi-scale feature information. The neck network structure constructed based on DF-BiFPN is shown below. Figure 4 shown.

[0069] Furthermore, in step S41, setting the hyperparameters for model training specifically includes:

[0070] The training rounds are 300, the patience value is 100, the batch size is 32, the optimizer is SGD, the image size is 640×640, and the number of threads is 4. The Mosaic online data augmentation method is used for training, and Mosaic enhancement is turned off 10 epochs before the end of training.

[0071] Furthermore, the step S42 specifically includes:

[0072] To evaluate the performance of the improved model, the experiment introduced evaluation metrics such as parameters (params), floating-point operations (FLOPs), recall, and mean average precision (mAP). Parameters represents the total number of learnable parameters in the model. FLOPs represents the total number of floating-point operations required for the model to complete a single forward pass, measuring the model's computational complexity.

[0073] The calculation formula for Recall is:

[0074]

[0075] Where TP represents the number of correctly detected positive samples, and FN represents the number of missed positive samples.

[0076] mAP is calculated by the following formula:

[0077]

[0078] In the formula, P represents the accuracy, FP represents the number of falsely detected negative samples, AP represents the average precision, N represents the total number of categories, and AP irepresents the average precision of the i-th category.

[0079] The test set is used to verify the detection effect of the optimal model for target detection, and the above evaluation indicators are combined to compare it with the original model YOLOv8n. The comparison results are as follows Figure 5 As shown in the figure, the three images on the left are the detection results generated by the original YOLOv8 model, and the three images on the right are the detection results generated by the improved DDG-YOLO model. The experimental results show that the improved DDG-YOLO model of the present invention has better detection effect.

[0080] Furthermore, to systematically evaluate the effectiveness of each improved module, we conducted multiple ablation experiments on the forest fire dataset using the YOLOv8n algorithm as the baseline model. The experiments tested the independent contributions and combined effectiveness of the three modules: DF-BiFPN (A), Dysample (B), and GN-SCDH (C). The results are shown in Table 1. Model performance was evaluated using the following metrics: parameter count, computational effort, recall, mAP50 (mean average precision at an IoU threshold of 0.50), and mAP50-95 (mean average precision at IoU thresholds between 0.50 and 0.95, with a step size of 0.05).

[0081] Table 1 Experimental results

[0082]

[0083] Note: √ means using this module, - means not using this module.

[0084] As shown in the experimental data in Table 1, the baseline model YOLOv8n (Test 1) has 3.01M parameters and 8.2GB of computation, achieving a recall of 66.6%, mAP50 of 73.7%, and mAP50-95 of 40.6%. Introducing the DF-BiFPN architecture (Test 2) reduces the number of parameters by 31% to 2.07M and the computation by 15% to 7.0GB, while improving mAP50 by 0.8 percentage points. Introducing the Dysample module (Test 3) improves recall and mAP50 by 0.7 and 0.6 percentage points, respectively, while maintaining model complexity. Introducing the GN-SCDH module (Test 4) reduces the number of parameters and computation by 21% and 20%, respectively, while maintaining stable accuracy. Combined experiments demonstrate that the combination of the DF-BiFPN architecture and the GN-SCDH module (Test 6) achieves optimal lightweighting, reducing the number of parameters by 44% to 1.70M and the computation by 29% to 5.8GB, while maintaining stable detection accuracy. The complete improved model DDG-YOLO (Experiment 8) reduces the number of parameters and computational complexity by 43% and 28% respectively, while the recall rate and mAP50 are further improved by 1.3 and 0.5 percentage points.

[0085] The improved DDG-YOLO model proposed in this paper achieves significant performance improvements in object detection by integrating the DF-BiFPN architecture, the Dysample module, and the GN-SCDH module. Compared to the baseline model YOLOv8n, this technical solution successfully reduces the model parameter count from 3.01M to 1.71M and the computational overhead from 8.2G to 5.9G. It also improves recall from 66.6% to 67.9% and mAP50 from 73.7% to 74.2%, achieving a dual optimization of model lightweighting and improved detection accuracy. In the specific application scenario of forest fire monitoring, this technical solution demonstrates significant advantages: its lightweight nature enables efficient model deployment on edge computing devices, significantly reducing hardware deployment costs. Furthermore, the improved detection accuracy and recall ensure accurate fire warnings, providing more reliable technical support for forest fire prevention. This simultaneous improvement in model efficiency, detection performance, and practical application value makes this invention a significant technological innovation in the field of object detection and offers broad market application prospects.

[0086] The above is a detailed description of the preferred embodiments of the present invention, but the scope of the present invention is not limited thereto. Those skilled in the art may make various adjustments or modifications without violating the core spirit of the present invention. All embodiments falling within the scope of the claims of this application are protected by the present invention.

Claims

1. A lightweight forest fire detection method based on DDG-YOLO, characterized by: include: S1: Construct an image dataset of forest fire scenes and preprocess the dataset; S2: Reconstruct the neck network structure based on the YOLOv8 network model: design a densely fused bidirectional feature pyramid network DF-BiFPN to optimize the neck network feature fusion, and use the dynamic upsampler Dysample to replace the traditional upsampling module Upsample; S3: Improved detection head structure: Introducing the shared convolutional detection head GN-SCDH with fusion group normalization to replace the original detection head and build an improved DDG-YOLO model; S4: Train the DDG-YOLO improved model, select the model with the best performance during training to detect forest fire images, and generate detection results.

2. A lightweight forest fire detection method based on DDG-YOLO according to claim 1, characterized in that: The specific implementation method of step S1 includes the following steps: S11: Collect an image dataset of forest fire scenes and add confusing scenes such as sunset glow, clouds, and fog to the dataset; S12: Use LabelImg software to label the dataset, and the output labeling file is a txt file in YOLO format; S13: The dataset is randomly divided into training set, test set and validation set in a ratio of 7:2:1; S14: Adjust the dataset image resolution to 640×640.

3. A lightweight forest fire detection method based on DDG-YOLO according to claim 1, characterized in that: The specific implementation method of step S2 includes the following steps: S21: Using the weighted feature fusion idea of the bidirectional feature pyramid network BiFPN, design the DF-BiFPN structure; S22: Dysample is used instead of the traditional Upsample to optimize the upsampling process of the feature map.

4. A lightweight forest fire detection method based on DDG-YOLO according to claim 1, characterized in that: The specific implementation method of step S3 includes the following steps: S31: Replace the group normalization of the convolution module with batch normalization to construct CGS convolution. The three feature maps output by the neck network are each passed through a 1×1 CGS convolution to unify the number of channels. S32: Input the result into the shared convolution layer for feature extraction. This layer contains two 3×3 CGS convolutions in series. S33: Process features through shared classification convolutional layers and regression convolutional layers, and introduce a Scale layer in the regression branch for feature scaling.

5. A lightweight forest fire detection method based on DDG-YOLO according to claim 1, characterized in that: The specific implementation method of step S4 includes the following steps: S41: Set the hyperparameters for model training, perform model training, and obtain the optimal model for target detection; S42: Use the test set to verify the detection effect of the optimal model for object detection, compare it with the original model based on the evaluation indicators, and verify the improvement effect of each module through ablation experiments; S43: Inputting the forest fire image to be detected into the optimal target detection model for detection to obtain the target detection result.

6. A lightweight forest fire detection method based on DDG-YOLO according to claim 3, characterized in that: The specific implementation method of step S21 is as follows: In traditional feature fusion methods, feature maps of different scales are weighted with fixed weights, resulting in insufficient feature adaptability and the inability to dynamically adjust the importance of features at each scale according to the characteristics of the detection target. BiFPN dynamically adjusts the importance of feature maps of different scales in the fusion process by introducing learnable weights, thereby more efficiently fusing multi-scale features. Its weighted feature fusion adopts a fast normalization method, as shown in formula (1). This method scales the weights to the range of 0 to 1 and uses the ReLU activation function to ensure that w i ≥0, and set ε=0.0001 to avoid numerical instability. Among them, O is the output feature, I i represents the input feature, ε is a numerical stability constant, and w i With w j Represents different weight learning parameters. While retaining the weighted feature fusion mechanism of BiFPN, DF-BiFPN employs a three-stage feature fusion strategy (top-down, bottom-up, and top-down) to construct a densely fused bidirectional feature pyramid network. Specifically, the architecture first performs channel-wise adjustments on the feature maps at the P3, P4, and P5 layers output by the backbone network through a 1×1 convolution, while simultaneously downsampling the P2 layer feature map through a 3×3 convolution. In the first top-down feature fusion stage, high-level features are progressively downsampled through 3×3 convolutions, and the downsampled results are weightedly fused with features at adjacent scales. Next, the network maintains the original bidirectional feature pyramid architecture of YOLOv8, and through a bottom-up and second top-down feature fusion process, ultimately constructing a pyramid structure that fully integrates feature information at multiple scales.

Images